chmm-lsim 1.0.0__tar.gz

chmm_lsim-1.0.0/.gitignore

@@ -0,0 +1,234 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[codz]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#   Usually these files are written by a python script from a template
+#   before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py.cover
+*.lcov
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+# Pipfile.lock
+
+# uv
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+# uv.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+# poetry.lock
+# poetry.toml
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#   pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
+#   https://pdm-project.org/en/latest/usage/project/#working-with-version-control
+# pdm.lock
+# pdm.toml
+.pdm-python
+.pdm-build/
+
+# pixi
+#   Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
+# pixi.lock
+#   Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
+#   in the .venv directory. It is recommended not to include this directory in version control.
+.pixi/*
+!.pixi/config.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule*
+celerybeat.pid
+
+# Redis
+*.rdb
+*.aof
+*.pid
+
+# RabbitMQ
+mnesia/
+rabbitmq/
+rabbitmq-data/
+
+# ActiveMQ
+activemq-data/
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.envrc
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs/Zensical documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#   JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#   be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#   and can be added to the global gitignore or merged into this file.  For a more nuclear
+#   option (not recommended) you can uncomment the following to ignore the entire idea folder.
+# .idea/
+
+# Abstra
+#   Abstra is an AI-powered process automation framework.
+#   Ignore directories containing user credentials, local state, and settings.
+#   Learn more at https://abstra.io/docs
+.abstra/
+
+# Visual Studio Code
+#   Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore that
+#   can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
+#   and can be added to the global gitignore or merged into this file. However, if you prefer, you
+#   could uncomment the following to ignore the entire vscode folder
+# .vscode/
+# Temporary file for partial code execution
+tempCodeRunnerFile.py
+
+# Ruff stuff:
+.ruff_cache/
+
+# PyPI configuration file
+.pypirc
+
+# Marimo
+marimo/_static/
+marimo/_lsp/
+__marimo__/
+
+# Streamlit
+.streamlit/secrets.toml
+
+# MATLAB
+*.asv
+*.m~
+*.mex*
+slprj/
+octave-workspace
+
+# benchmark outputs
+benchmarks/*.csv
+
+# patch backups
+*.orig
+*.rej

chmm_lsim-1.0.0/CHANGELOG.md

@@ -0,0 +1,24 @@
+# Changelog
+
+All notable changes to chmm-lsim-x are recorded here. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and versions follow
+[Semantic Versioning](https://semver.org/spec/v2.0.0.html):
+
+- **patch** (1.0.x): bug fixes only; no change to the public API;
+- **minor** (1.x.0): new models, functions, options or backends; existing code keeps working unchanged;
+- **major** (x.0.0): the only releases that may remove or change public API, and only after at least one minor release in which the old form still works and issues a `DeprecationWarning` (MATLAB: a `warning`).
+
+The public API is every name exported by `chmm_lsim` (its `__all__`) and by `chmm_lsim.discriminative`, and every MATLAB function in `matlab/` (outside `tests/`), with its documented arguments and outputs. Everything else (Python submodules such as `chmm_lsim.models.lsim.em`) is internal and may be reorganized in any release. Models saved with `save_lsim_params` stay loadable by later versions.
+
+## [Unreleased]
+
+## [1.0.0]
+
+First public release, accompanying the POISE paper.
+
+### Added
+- **LSIM inference** with three backward recursions: POISE (the default), SED and fast/SVI, in MATLAB (`forward_backward_lsim`, `lsim_fb_constants`) and Python (`forward_backward`, `FbConstants`); NumPy, numba and PyTorch (CPU/CUDA, batched) backends in Python.
+- **LSIM learning:** maximum-likelihood EM (`em_lsim`), mini-batch training for large datasets (`em_lsim_minibatch_train`), parameter initialization, simulation, approximate Viterbi decoding (`viterbi_lsim`), exact inference for small models, model save/load.
+- **Full CHMM:** EM (`em_chmm`), exact forward-backward (`forward_backward_chmm`), exact Viterbi (`viterbi_chmm`) and simulation, in MATLAB and Python, with outputs compatible with `em_lsim` (MATLAB: the same output order without the coupling matrix; Python: the same `EmResult`).
+- **Discriminative (MMI) learning** of class-conditional LSIMs: `em_lsim_discriminative`, `lsim_discriminative_predict` (MATLAB and Python).
+- Runtime benchmark of the backward recursions (`benchmarks/`), demos in MATLAB and Python, tests checked against MATLAB reference outputs.

chmm_lsim-1.0.0/LICENSE

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2026, Sajjad Karimi
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

chmm_lsim-1.0.0/PKG-INFO

@@ -0,0 +1,235 @@
+Metadata-Version: 2.5
+Name: chmm-lsim
+Version: 1.0.0
+Summary: Inference and learning for coupled hidden Markov models and Latent Structure Influence Models (LSIMs): POISE, SED and SVI forward-backward, EM, and discriminative (MMI) training
+Project-URL: Homepage, https://github.com/sajjadkarimi91/chmm-lsim-x
+Project-URL: Repository, https://github.com/sajjadkarimi91/chmm-lsim-x
+Project-URL: Issues, https://github.com/sajjadkarimi91/chmm-lsim-x/issues
+Project-URL: Changelog, https://github.com/sajjadkarimi91/chmm-lsim-x/blob/main/CHANGELOG.md
+Author: Sajjad Karimi
+License-Expression: BSD-3-Clause
+License-File: LICENSE
+Keywords: chmm,coupled-hmm,em-algorithm,forward-backward,hidden-markov-model,influence-model,lsim,multichannel,time-series
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Scientific/Engineering :: Information Analysis
+Requires-Python: >=3.10
+Requires-Dist: numpy>=1.26
+Requires-Dist: scipy>=1.11
+Provides-Extra: all
+Requires-Dist: numba>=0.60; extra == 'all'
+Requires-Dist: scikit-learn>=1.3; extra == 'all'
+Requires-Dist: threadpoolctl>=3; extra == 'all'
+Requires-Dist: torch>=2.4; extra == 'all'
+Provides-Extra: dev
+Requires-Dist: numba>=0.60; extra == 'dev'
+Requires-Dist: pytest>=8; extra == 'dev'
+Requires-Dist: scikit-learn>=1.3; extra == 'dev'
+Requires-Dist: threadpoolctl>=3; extra == 'dev'
+Requires-Dist: torch>=2.4; extra == 'dev'
+Provides-Extra: init
+Requires-Dist: scikit-learn>=1.3; extra == 'init'
+Requires-Dist: threadpoolctl>=3; extra == 'init'
+Provides-Extra: numba
+Requires-Dist: numba>=0.60; extra == 'numba'
+Provides-Extra: torch
+Requires-Dist: torch>=2.4; extra == 'torch'
+Description-Content-Type: text/markdown
+
+# chmm-lsim-x
+
+Inference and learning for **coupled hidden Markov models (CHMMs)** and **Latent Structure Influence Models (LSIMs)**, in **MATLAB** and **Python**.
+
+An LSIM models C interacting channels (for example EEG electrodes or sensors). Each channel has its own hidden Markov chain, and the next state of channel ζ is drawn from a mixture of pairwise transitions from every channel c, weighted by coupling strengths θ(c, ζ):
+
+P(x<sub>t</sub><sup>ζ</sup> | x<sub>t−1</sub>) = Σ<sub>c</sub> θ(c, ζ) · A<sup>c,ζ</sup>(x<sub>t−1</sub><sup>c</sup>, x<sub>t</sub><sup>ζ</sup>)
+
+This keeps the model size at O((Σ M)²) instead of the O(Π M) joint state space of a full CHMM.
+
+## What is included
+
+| | MATLAB (`matlab/`) | Python (`chmm_lsim`) |
+|---|---|---|
+| Forward-backward inference: **POISE** (default), **SED**, **fast/SVI** | `forward_backward_lsim` + `lsim_fb_constants` | `forward_backward`, `FbConstants` |
+| Exact inference for small models (equivalent joint HMM) | `forward_backward_chmm`, `viterbi_chmm`, `em_lsim_exact` | `forward_backward_chmm`, `viterbi_chmm`, `exact_posterior`, `viterbi_exact` |
+| Maximum-likelihood EM learning | `em_lsim`, `em_lsim_poise`, `em_lsim_fast` | `em_lsim`, `em_lsim_minibatch_train` |
+| Full CHMM (joint-state transitions): EM, exact inference, exact Viterbi, simulation | `em_chmm`, `forward_backward_chmm`, `viterbi_chmm`, `generate_chmm_time_series` | `em_chmm`, `forward_backward_chmm`, `viterbi_chmm`, `simulate_chmm`, `initialize_chmm_params`, `lsim_to_chmm` |
+| Discriminative learning (**MMI**) of class-conditional LSIMs | `em_lsim_discriminative`, `lsim_discriminative_predict` | `em_lsim_discriminative`, `lsim_discriminative_predict` |
+| LSIM simulation, approximate Viterbi decoding, model save/load | `generate_chmm_time_series`, `viterbi_lsim` | `simulate_lsim`, `viterbi_lsim`, `save_lsim_params` |
+| Execution | CPU | NumPy, numba, PyTorch (CPU or CUDA, batched) |
+
+All observations may contain missing samples (`NaN`). Emissions are Gaussian mixtures with diagonal or full covariance.
+
+`em_lsim` and `em_chmm` share the same options, initialization, emission update, convergence rule and outputs: in MATLAB, `em_chmm` returns `em_lsim`'s outputs in the same order without the coupling matrix (`[pi_0, transitions, gmm, AIC, log_likelihood, BIC, pi_steady]`), and in Python both return the same `EmResult`. `viterbi_lsim` is a fast approximate per-channel decoder for LSIMs; `viterbi_chmm` decodes a CHMM exactly on the joint state space (in MATLAB, `viterbi_chmm` called with LSIM parameters still runs `viterbi_lsim`).
+
+### Choosing an inference recursion
+
+All three recursions share the same forward pass, so they give the same approximate log-likelihood at fixed parameters. They differ in the backward pass:
+
+- **POISE** (default): weighs each channel's outgoing influence by its total outgoing coupling strength. It needs no pairwise (S×S) message, is the fastest, and gives the most accurate smoothed marginals in our simulations, especially for strongly coupled channels.
+- **SED**: builds a pairwise conditional message and solves a small linear system per channel at every step. Slower, and uses more memory.
+- **fast/SVI**: fixed coupling weights raised to a fixed power. As fast as POISE.
+
+Use exact inference (Cartesian product of channel states) only for small models; its cost grows with the product of the state counts.
+
+## Installation
+
+**MATLAB** (tested with R2024b; the GMM initialization in `em_lsim` uses the Statistics and Machine Learning Toolbox):
+
+```matlab
+run('<path to chmm-lsim-x>/matlab/setup_chmm_lsim.m')   % or: addpath(genpath('<path to chmm-lsim-x>/matlab'))
+```
+
+**Python** (3.10 or later):
+
+```bash
+pip install "chmm-lsim[all] @ git+https://github.com/sajjadkarimi91/chmm-lsim-x"
+```
+
+Optional extras: `init` (scikit-learn, for `initialize_lsim_params`), `numba` (numba CPU backend), `torch` (PyTorch CPU/CUDA backend), `all` (all three). The core needs only NumPy and SciPy. For CUDA, install a CUDA build of PyTorch first (see pytorch.org).
+
+## Quick start: MATLAB
+
+Observations are a C×1 cell array; `observations{c}` is a d(c)×T matrix (`NaN` = missing). For several sequences, use a C×N cell array.
+
+```matlab
+% EM learning: 4 hidden states and 1 Gaussian per channel, 100 iterations, POISE E-step
+extra = struct('plot', 0, 'sigma_diag', 1);
+[pi_0, coupling, transitions, gmm_para, AIC, log_lik, BIC] = ...
+    em_lsim(observations, 4*ones(1,C), ones(1,C), 100, extra);
+
+% Smoothed marginals P(x_t^c | all data); marginals{c} is M(c)xT
+[log_lik, ~, ~, marginals] = forward_backward_lsim(pi_0, coupling, transitions, gmm_para, observations);
+
+% Choose the recursion explicitly
+fbc = lsim_fb_constants('sed');           % 'poise' (default), 'sed' or 'fast'
+[~, ~, ~, marginals_sed] = forward_backward_lsim(pi_0, coupling, transitions, gmm_para, observations, fbc);
+extra.fb_constants = fbc;                 % the same option selects the E-step of em_lsim
+```
+
+**MMI training** of one LSIM per class, starting from EM models:
+
+```matlab
+% class_obs{k}: C x N_k cell of class-k training sequences; val_obs: same format
+for k = 1:numel(class_obs)                % generative start: EM per class
+    [pi_0, theta, A, gmm] = em_lsim(class_obs{k}, M, K, 30, extra);
+    init_models{k} = struct('pi_0_lsim', {pi_0}, 'coupling_tetha_IM', theta, ...
+        'transition_matrices_IM_cell', {A}, 'gmm_para_lsim', {gmm});
+end
+[models, history] = em_lsim_discriminative(class_obs, M, K, [], ...
+    struct('objective', 'mmi', 'validation_observations', {val_obs}), init_models);
+[scores, predicted] = lsim_discriminative_predict(models, test_obs);   % test_obs: C x N cell
+```
+
+**Full CHMM** (small models; parameters have the same cell format as `em_lsim`'s, with one prod(M) x M(c) transition table per channel):
+
+```matlab
+[pi_0, transitions, gmm, AIC, log_lik, BIC] = em_chmm(observations, M, K, 100, extra);
+[log_lik, ~, ~, marginals] = forward_backward_chmm(pi_0, transitions, gmm, observations);   % exact
+[log_prob, states] = viterbi_chmm(pi_0, transitions, gmm, observations);                  % exact
+```
+
+Demos: `matlab/demos/demo_run_lsim.m` (POISE, SED and fast vs exact inference, then EM), `demo_run_chmm.m` (CHMM vs LSIM) and `demo_discriminative_lsim.m` (EM vs MMI classification).
+
+## Quick start: Python
+
+Observations of one sequence are a (D, T) array with the channels stacked along D (`NaN` = missing); a dataset is a list of such arrays.
+
+```python
+from chmm_lsim import (LSIMStructure, FbConstants, initialize_lsim_params, em_lsim,
+                       forward_backward, viterbi_lsim)
+
+structure = LSIMStructure(channel_state_num=(4, 4, 4), channel_obs_dim=(1, 2, 1),
+                          num_gmm_component=(1, 1, 1))
+params0 = initialize_lsim_params(structure, Y_list, sigma_diag=True, seed=0)   # needs [init]
+result = em_lsim(structure, params0, Y_list, max_iterations=100)                # POISE E-step
+post = forward_backward(structure, result.params, Y_list[0])                    # post.gamma: (S, T)
+post_sed = forward_backward(structure, result.params, Y_list[0], FbConstants(variant="sed"))
+states = viterbi_lsim(structure, result.params, Y_list[0]).states
+```
+
+Large datasets and GPUs:
+
+```python
+from chmm_lsim import em_lsim_minibatch_train, forward_backward_batch
+
+loglik = forward_backward_batch(structure, result.params, Y_list, mode="loglik",
+                                device="cuda", dtype="float32")                 # batched scoring
+fit = em_lsim_minibatch_train(structure, params0, Y_list, max_iterations=20, num_minibatches=10,
+                              backend="torch", backend_options={"device": "cuda", "dtype": "float64"})
+```
+
+MMI training:
+
+```python
+from chmm_lsim import em_lsim_discriminative, lsim_discriminative_predict
+
+res = em_lsim_discriminative((Y_train, y_train), init_models,   # one LSIMParams per class
+                             validation=(Y_val, y_val), objective="mmi")
+scores, predicted, normalized = lsim_discriminative_predict(res.models, Y_test)
+```
+
+Full CHMM (exact, for small models; `em_chmm` returns the same `EmResult` as `em_lsim`):
+
+```python
+from chmm_lsim import initialize_chmm_params, em_chmm, forward_backward_chmm, viterbi_chmm
+
+fit = em_chmm(structure, initialize_chmm_params(structure, Y_list, seed=0), Y_list, max_iterations=100)
+post = forward_backward_chmm(structure, fit.params, Y_list[0])      # post.gamma: (S, T)
+states = viterbi_chmm(structure, fit.params, Y_list[0]).states       # exact joint decoding
+```
+
+Demos: `examples/demo_run_lsim.py` and `examples/demo_run_chmm.py`.
+
+## Benchmarks
+
+`benchmarks/` reproduces the runtime and memory comparison of the POISE, SED and fast/SVI backward recursions on CPU and GPU, and the full E-step on 100,000 sequences. See `benchmarks/README.md`.
+
+## Repository layout
+
+```
+matlab/
+  setup_chmm_lsim.m  adds the toolbox folders to the MATLAB path
+  lsim/              LSIM inference (POISE, SED, fast/SVI), EM, exact EM, constants, approximate Viterbi
+  lsim/backward/     one file per LSIM backward recursion
+  chmm/              full CHMM: EM, exact forward-backward, exact Viterbi, simulation
+  emissions/         Gaussian-mixture densities and sampling
+  common/            parameter helpers
+  discriminative/    MMI training, prediction, metrics (+ tests/)
+  demos/
+src/chmm_lsim/       Python package
+  core/              structure, validation, constants, serialization, joint-state utilities
+  emissions/         emission models (Gaussian mixtures) and their registry
+  models/lsim/       LSIM: inference (POISE, SED, fast), EM, initialization, simulation, Viterbi
+  models/chmm/       full CHMM
+  learning/          shared EM loop, mini-batch training, discriminative (MMI)
+  backends/          NumPy, numba, PyTorch
+tests/               Python tests (pytest), with MATLAB reference fixtures
+examples/            Python demos
+benchmarks/          runtime benchmark
+docs/EXTENDING.md    how to add inference recursions, emission models and model families
+```
+
+## Versions and compatibility
+
+Releases follow [semantic versioning](https://semver.org): new models and features arrive in minor releases without breaking existing code, and anything public that is ever renamed or removed keeps working with a deprecation warning for at least one minor release first. The public API is everything importable from `chmm_lsim` (and `chmm_lsim.discriminative`) and every MATLAB function; Python submodules are internal, so import from `chmm_lsim` directly. Changes are listed in [`CHANGELOG.md`](CHANGELOG.md); [`docs/EXTENDING.md`](docs/EXTENDING.md) explains how new models are added.
+
+## Tests
+
+```bash
+pip install -e ".[dev]"
+pytest
+```
+
+The Python implementation is checked against MATLAB reference outputs (`tests/fixtures/`). MATLAB tests of the MMI trainer: `runtests('matlab/discriminative/tests/TestMMI.m')`.
+
+## License
+
+BSD 3-Clause. See `LICENSE`.